search

WoodwindHu / RangeLDM

[ECCV 2024] Official implementation of "RangeLDM: Fast Realistic LiDAR Point Cloud Generation"
MIT License
17 stars 0 forks source link

Correct vae_checkpoint to run the ldm nuScenes training #3

Open JackieHuJunyi opened 1 week ago

JackieHuJunyi commented 1 week ago

The vae_checkpoint you provided is diffusion_pytorch_model.safetensors, not a .ckpt or .pth file. But in ldm/train_unconditional.py, vae_checkpoint = torch.load(args.vae_checkpoint, map_location='cpu')['state_dict']. How is this done? If I use vae_checkpoint = load_file(args.vae_checkpoint) and vae_checkpoint = {k: v.to("cpu") for k, v in vae_checkpoint.items()} instead, the problem of "NotImplementedError: Cannot copy out of meta tensor; no data!" cannot be solved.

JackieHuJunyi commented 1 week ago

For Example: Parameter 'encoder.mid_block.resnets.0.norm1.weight' is on meta device.